A Pattern Languagefor Agentic Systems
PatternsCompositionsMethodologiesTrainingsSearchUse →
  1. Catalog
  2. Compositions
  3. Recipes
  4. Safety Hardening

Recipe · Recipes

Safety Hardening

Type: recipe · Status in practice: mature

❖ ❖ ❖

Description. The minimum set of constraints to put around any production agent before it touches the world: budgets, gates, charters, kill-switches, approvals.

Patterns this recipe composes —

  • ★★Step Budget
    core
  • ★Kill Switch
    core
  • ★★Approval Queue
    core
  • ★Constitutional Charter
    core
  • ★Tenant-Scoped Tool Binding
    core

    Enforces tenant scope below the prompt so a misled model cannot widen data access.

  • ★★Cost Gating
    hardening
  • ★★Rate Limiting
    hardening
  • ★Sandbox Escape Monitoring
    hardening
  • ★★Refusal
    hardening
  • ★★Compensating Action
    hardening
  • ★Risk-Tiered Action Autonomy
    hardening

    Holds material actions for a human approver distinct from the initiator.

  • ★Tool Output Poisoning Defense
    optional

Neighbourhood

Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.

Instantiated by (9)

  • full-code · framework
    Cline
    Full-Code
  • full-code · framework
    Codex CLI
    Full-Code
  • full-code · framework
    Continue
    Full-Code
  • app · closed product
    Devin
    App
  • app · closed product
    GitHub Copilot Coding Agent
    App
  • full-code · framework
    JetBrains Junie
    Full-Code
  • full-code · framework
    OpenHands
    Full-Code
  • full-code · framework
    Plandex
    Full-Code
  • full-code · framework
    Windsurf
    Full-Code

Provenance

  • Last updated: 2026-06-10
↑ Recipesall compositions

Based on the open Agent Patterns Catalog — CC BY 4.0 · support & legal · ·